Replace cumulative Hindsight transcript snapshots

[discolotus]

Summary

• switch modern Hindsight retain batches from update_mode=append on the session document to update_mode=replace on the per-process document
• avoid repeatedly appending cumulative transcript snapshots, which duplicated earlier turns and grew documents indefinitely
• keep legacy API behavior unchanged by omitting update_mode when unsupported

Rationale

Hermes builds each auto-retain payload from the full in-process session buffer. Appending those cumulative snapshots duplicates earlier turns on every retain. Replacing the per-process snapshot preserves the current process lifecycle transcript without risking a restart overwriting a prior process lifecycle.

Testing

• /Users/tleo/.hermes/hermes-agent/venv/bin/python -m pytest tests/plugins/memory/test_hindsight_provider.py -q
• /Users/tleo/.hermes/hermes-agent/venv/bin/python -m py_compile plugins/memory/hindsight/__init__.py run_agent.py gateway/run.py tests/plugins/memory/test_hindsight_provider.py

[teknium1]

This looks implemented on current main by a later Hindsight fix. Automated hermes-sweeper review.

Evidence:

• plugins/memory/hindsight/__init__.py:1473 now sends only turns accumulated since _last_retained_turn_count when the API supports update_mode='append', instead of re-sending the cumulative transcript snapshot on every retain.
• plugins/memory/hindsight/__init__.py:1533 advances that append watermark after queueing the retain, so later retains do not re-ship turns already handed to the writer.
• tests/plugins/memory/test_hindsight_provider.py:805 covers the regression: the second append retain excludes the earlier turns and includes only the new turn batch.
• The fix landed in 09d66037f8f7bc5bd879ed8128273fb6780a009f (fix(hindsight): send only new-turn delta on append retains instead of whole session (#40605)).

Main uses a different implementation than this PR proposed: it keeps the stable session-scoped append document and sends deltas, rather than switching modern APIs to per-process replace. That still resolves the reported duplication/unbounded-growth behavior.